10 million downloads: Reported privacy leaks hit Meta's fast-growing Muse agent
A Reuters investigation based on internal Meta posts describes how the company's fast-growing AI agent has reportedly mixed up conversations across accounts and exposed private data. Meta says the episodes are either isolated or hallucinations — but could not prove it.
Meta employees have reportedly reported internally that the company's AI agent Muse is mixing up conversations from different user accounts and exposing private information that does not belong to the user. That is the core finding of a new Reuters investigation based on internal company posts, covered by both TechCrunch and ZDNET on 19 November 2026. Everything in this story is secondhand reporting on the Reuters report: the claims are reported, not established facts, and Meta's own response — that the episodes are either isolated coincidences or model hallucinations — is, according to Reuters, unverified.
The specific episodes
In several of the cases, employees reportedly told colleagues that Muse suddenly displayed messages they had never written or received: childcare requests they had never made, career advice they had never asked for, product orders they had never placed, and bank balances belonging to someone else entirely.
The most concrete example is an employee who, according to TechCrunch, which cites Reuters, told colleagues that Muse, in the middle of an entirely ordinary chat about weather and tides, sent an additional message with two bank updates: "BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57" — that is, the agent reportedly stated the balances of the user's Capital One and Charles Schwab accounts to another person, in a context where that information did not belong.
Another case concerns end-to-end encrypted communication: an employee who used Muse on their personal phone reportedly received a reminder from the agent to reply to a message from their sister in Signal — an app Muse had not been given access to. And in a third case, Muse reportedly booked a non-refundable Airbnb reservation for an employee who had never asked the agent to do so, according to ZDNET's account of the Reuters investigation.
How it appears to happen
Reuters and the think tank AI Now Institute have tested Muse themselves. On an Android phone and a Mac, they found that the agent uses permissions granted during the bot's initial setup to pull in message text — "leveraging those authorizations in ways a user might not anticipate," as ZDNET renders the finding. The same test did not work on an iPhone.
This suggests the mechanism is not that Muse "clones" conversations between accounts in the backend, but that the agent is granted read access to notifications and messages on the device during setup — and can then pull in content it should not use, or use it in the wrong context. The Signal episode points in the same direction: an agent with access to notification content can pick up text from apps the user never intended to be part of the agent's scope of work.
It is worth emphasizing what the test actually showed: a successful reconstruction on Android and Mac, a failed one on iPhone. It establishes a plausible mechanism, not a documented scope.
Meta's response — and what remains unresolved
To Reuters, Meta has explained the episodes as either isolated individual cases or hallucinations from the model — that is, the agent may have invented the balances rather than leaked them. Reuters could not confirm this explanation. In practice, the matter therefore remains unresolved: either leaks of real user data, or a model making up highly specific, apparently private information — both would be serious for an agent with this kind of access, but in different ways.
Meta did not immediately respond to TechCrunch's request for comment, so there is no verified comment from the company in the TechCrunch piece; Meta's explanation is known only via ZDNET's account of Reuters.
The context: 10 million downloads in two months
Muse launched in September and has, according to TechCrunch, passed 10 million downloads, making it one of the fastest-growing apps ever measured. This is not a niche product: reported errors of this kind concern an agent with an exceptionally fast-growing user base, where permissions are granted at setup and rarely reviewed again.
Nor is this the first security problem for Muse. As early as the launch month, a security expert discovered a serious zero-day vulnerability that made the app hijackable, according to ZDNET — a case of a different type than what is now being reported, but part of the same short history of early weaknesses.
What users can do
The concrete advice that follows from the findings is to review which permissions Muse has on your device — particularly notification and messaging permissions on Android, where the Reuters and AI Now Institute test showed that setup permissions were used to pull in message text. On iPhone, the same test method did not work, but permissions the app does not need can be restricted regardless of platform.
Open questions
The most pressing unanswered question is scope: how often does conversation mixing occur among the more than 10 million users, and does it affect only Meta employees or external users as well? It is also unclear why the reported cases so far have been tied to the company's own employees — could this be due to how they use the agent, or is it simply where internal reporting takes place? And finally: if Meta's "hallucination" explanation were to hold, another serious question remains — why an agent with access to such sensitive contexts can generate convincing fake bank balances and private messages. None of these questions are answered by what is currently known.

